| Objective: The early diagnosis of liver fibrosis is of great significance in preventing liver cirrhosis and liver cancer.As gold standard for staging liver fibrosis,liver biopsy is an invasive procedure that carries the risk of serious complications.Non-invasive diagnosis methods based on big data and deep learning models are gradually being widely applied in clinical work.The aim of this study was to evaluate the performance of the residual neural network(ResNet)for staging liver fibrosis using plain CT images.Methods: This retrospective study involved 347 patients subjected to liver plain CT scanning and liver biopsy,including 213 males and 134 females.The liver plain CT examination was completed within three months before or after the puncture.The precise pathological staging of liver fibrosis patients is divided into five categories: no fibrosis(F0),portal fibrosis without septa(F1),portal fibrosis with few septa(F2),numerous septa without cirrhosis(F3)and liver cirrhosis(F4).In clinical pathological staging diagnosis,F0 and F1 are defined as no significant liver fibrosis,?F2 is significant liver fibrosis,?F3 is advanced liver fibrosis,and F4 is liver cirrhosis.For each patient,we selected three axial images adjacent to the puncture location.After processing and enhancements,these images were used as input data for the ResNet model.First of all,we established a two-class ResNet model.The output labels of the two-class ResNet model were 0 and 1,and they were composed of the following conditions: ?1 0: F0-3,1: cirrhosis(F4)?2 0: F0-2,1: advanced liver fibrosis(F3-4)?3 0:F0-1,1: significant liver fibrosis(F2-4)?4 0: F0,1: mild liver fibrosis(F1-4).The receiver operating characteristic curve and accuracy,sensitivity,specificity were used to evaluate the ability of the two-class ResNet model to diagnose liver cirrhosis(F4 vs.F0-3),advanced liver fibrosis(F3-4 vs.F0-2),significant liver fibrosis(F2-4 vs.F0-1),and presence of liver fibrosis(F1-4 vs.F0).Using similar modeling methods,two five-class ResNet models involved 347 patients for precise staging were established.First,we established a five-class ResNet model just based on CT plain scan images.Besides,we collected part of the patient's clinical information,including age,gender and years of liver abnormality.We merged the clinical features and CT plain scan image features in the input layer to obtain the five-class ResNet mixed model.The output labels of these two five-class accurate staging models are 0,1,2,3 or 4,which correspond to F0,F1,F2,F3,or F4 respectively.According to the output results,two five-class confusion matrixes of the two models were obtained.Kappa coefficient,accuracy,recall rate and precision rate were used to evaluate the performance of five-class ResNet model and five-class ResNet mixed model to diagnose liver fibrosis in accurate staging F0,F1,F2,F3 and F4.The ability of the five-class ResNet model and the five-class ResNet mixed model to assess the precise staging of liver fibrosis was compared.In addition,we used 318 chronic hepatitis B or C patients with plain CT images and clinical information to train the five-class ResNet mixed model.The model was optimized from three perspectives: increasing the image diversity,controlling the cause of liver fibrosis,and balancing the proportion of each liver fibrosis stage.And through the five-class confusion matrix of the model and related kappa coefficient values,the ability of the five-class ResNet mixed model to diagnose liver fibrosis for precise staging F0,F1,F2,F3 and F4 in patients with chronic hepatitis B or C was evaluated.The models used a five-fold cross-validation method.Approximately 80% images of the total sample size were used for training the models,the other 20% were used for testing the trained network,with the liver biopsy pathology results as gold standard.The proportion of patients in each fibrosis stage was equal for training and test groups.The final result was the mean of the five-fold cross-validation in the test group.Results:(1)In two-class staging of liver fibrosis in patients with chronic liver disease,The area under curve(AUC)of ResNet model for assessing liver cirrhosis(F4),advanced liver fibrosis(F3 or higher),significant liver fibrosis(F2 or higher),and presence of fibrosis(F1 or higher)was 0.97,0.94,0.90,and 0.91,respectively.(2)In five-class staging of liver fibrosis in patients with chronic liver disease,the kappa coefficient calculated from the confusion matrix of the five-class ResNet model to diagnose liver fibrosis in patients with chronic liver disease was 0.566.That is,the five-class ResNet model had moderate consistency with the gold standard pathology.The five-class ResNet model had a recall rate of 14.7%,5%,60%,23.5% and 86.3%for F0,F1,F2,F3 and F4 stages,respectively.While the kappa coefficient calculated from the confusion matrix of the ResNet five-class mixed model to diagnose liver fibrosis in patients with chronic liver disease was 0.63.That is,the ResNet five-class mixed model had high consistency with the gold standard pathology.The recall rate of the ResNet mixed model for F0,F1,F2,F3 and F4 was 26.7%,6.9%,47.2%,11.8% and 84.5%,respectively.The diagnostic results between the five-class ResNet model and the five-class ResNet mixed model were significantly different(p<0.05).(3)After optimization from the three perspectives of increasing image diversity,controlling the etiology and balancing the proportion of each liver fibrosis stage,the five-class ResNet mixed model had a recall rate of 55.4%,13.6%,36.4%,18.3% and76.5% for F0,F1,F2,F3,and F4,respectively.The kappa coefficient calculated based on the five-class confusion matrix of the optimized five-class ResNet mixed model was0.52.That is,in staging liver fibrosis in patients with chronic hepatitis B or C,the five-class ResNet mixed model was moderately consistent with the gold standard pathology.Conclusions:(1)The two-class ResNet model based on plain CT images exhibited high diagnostic efficiency for liver fibrosis staging.It can be used as a non-invasive imaging method to assist in the diagnosis of liver cirrhosis,advanced liver fibrosis,significant liver fibrosis,and presence of liver fibrosis.(2)In the precise staging of liver fibrosis,the five-class ResNet model based on CT plain scan images and the five-class ResNet mixed model based on CT plain scan images and clinical information were of higher value in the diagnosis of liver cirrhosis.However,their ability of precise staging F0,F1,F2 and F3 needed to be improved.Compared with the five-class ResNet model,the five-class ResNet mixed model adding clinical information improved its ability in accurate staging of liver fibrosis.(3)The optimized five-class ResNet mixed model could assist in the diagnosis of cirrhosis in patients with chronic hepatitis B or C.Although the recall rate and accuracy rate of the model for F0,F1,F2,and F3 stages had been slightly improved,its ability to diagnose liver fibrosis in non-cirrhosis patients still needed to be improved.In short,the deep residual network ResNet model based on CT plain scan images can assist in the diagnosis of liver cirrhosis.For patients with non-cirrhosis,it can assess the trend of fibrosis,but it is not enough to diagnose liver fibrosis accurately. |
PDF Full Text Request